MOE models and -mlock

[ZorXL]

If a MOE model is bigger than available RAM, then -mlock doesn't work. I wonder is somehow possible using mlock or perhaps using -ot (tensor overriding) to "mlock" just shared experts and attention, for example.